I am currently working for having A.I. workflow for generating projects. There are two things that need to be tackled for that: 1) hallucinations; and 2) old codes. I have met several patterns to tackle those two and just need to formulate the patterns into a general flow. Thankfully, a video of building fullstack app using A.I. from Raf Dev channel inspired me to make the formulation. He used multiple A.I. engines to his need: He used Google Gemini to onboard his ideas and summarized it. He forked a boilerplate of NextJS project into a new project. Then, he added the summary from Google Gemini into a file for context. He used QWEN Code, a QWEN3 code engine that is a fork from Gemini CLI, to build the project. When working with an existing project, A.I. will take the whole project as a context. It also means that it will stick with the version used by the libraries. It also will try to update its knowledge based on the common pattern in the project. Most of the time, it will use the c...
I am learning how to do vibe coding right. Vibe coding was a term coined by Andrej Karpathy for doing coding using LLM. And the world splitted with acknowledging it and flatly denying it. There is a compelling reason why vibe coding is dangerous: LLM trained on a snapshot knowledge. Meaning, the updated knowledge might get passed out. This makes LLM produces old syntax that are no longer true. This is dangerous for building web where we are using the volatile Javascript libraries. LLM have the tendency to halucinate. This means it would generate codes that doesn't work. This is actually the problem with the way LLM work previously. It only has a set of knowledge built on. The RAG also might not correct. They have fixed this with the new LLM specialized code. This was started by Antropic's Claude Code . And then, folks at Google releases the same with Gemini CLI . After that, Alibaba forks the Gemini CLI and built QWEN Code . These are built either run in the project directory...